Post on 21-Dec-2015
Replication & Consistency
EECE 411: Design of Distributed Software Applications
Logistics
Project P01 Non-blocking IO lecture: Nov 4th. P02 deadline on NEXT Wednesday (November 17th)
Next quiz Tuesday: November 16th
EECE 411: Design of Distributed Software Applications
Problem Technique AdvantagePartitioning Consistent hashing Incremental scalability
High availability for writes
Eventual consistency Quorum protocol
Vector clocks with reconciliation during
reads
High availability
Handling temporary failures
‘Sloppy’ quorum protocol and hinted
handoff
Provides high availability and
durability guarantee when some of the replicas are not
available.
Recovering from permanent failures
Anti-entropy using Merkle trees
Synchronizes divergent replicas in the background.
Membership and failure detection
Gossip-based membership protocol
Preserves symmetry and avoids having a
centralized registry for storing membership and node liveness
information.
Last time: Amazon’s Dynamo service
EECE 411: Design of Distributed Software Applications
Replication: Creating and using multiple copies of data (or services)
Why replicate? Improve system reliability
Prevent data loss i.e. provide data durability
Increase data availability Note: availability ≠ durability
Increase confidence: e.g. deal with byzantine failures
Improve performance Scaling throughput Reduce access times
EECE 411: Design of Distributed Software Applications
Issue 1. Dealing with data changes Consistency models
What is the semantic the system implements (luckily) applications do not always require strict
consistency Consistency protocols
How to implement the semantic agreed upon?
Issue 2. Replica management How many replicas? Where to place them? When to get rid of them?
Issue 3. Redirection/Routing Which replica should clients use?
What are the issues?
EECE 411: Design of Distributed Software Applications
Performance and Scalability
To keep replicas consistent, we generally need to ensure that all conflicting operations are done in the same order everywhere
Conflicting operations: From the world of transactions: Read–write conflict: a read operation and a write
operation act concurrently Write–write conflict: two concurrent write operations
Problem: Guaranteeing global ordering on conflicting operations may be costly, reducing scalability
Solution: Weaken consistency requirements so that hopefully global synchronization can be avoided
Scalability TENSION Management overheads
EECE 411: Design of Distributed Software Applications
Client‘s view of the data-store
Ideally: black box -- complete ‘transparency’ over how data is stored and managed
EECE 411: Design of Distributed Software Applications
Management’ system view on data store
Management system: Controls the allocated resources and aims to provide transparency
EECE 411: Design of Distributed Software Applications
Consistency model
Management system: Controls the allocated resources and aims to provide transparency
Consistency model: Contract between the data store and the clients: The data store specifies the results of read and write operations in the presence of concurrent operations.
EECE 411: Design of Distributed Software Applications
Roadmap for next few classes
Consistency models: contracts between the data store and the
clients that specify the results of read and write operations are in the presence of concurrency.
Protocols To manage update propagation To manage replicas:
creation, placement, deletion To assign client requests to replicas
EECE 411: Design of Distributed Software Applications
Consistency models
Data centric: Assume a global, data-store view (i.e., across all clients) Continuous consistency
Limit the deviation between replicas Models based on uniform ordering of
operations Constraints on operation ordering at the data-store
level Client centric
Assume client-independentviews of the datastore
Constraints on operation ordering for each client independently
EECE 411: Design of Distributed Software Applications
Notations
For operations on a Data Store: Read: Ri(x) a -- client i reads a from
location x Write: Wi(x) b -- client i writes b at location x
P1: W(x)1
P2: R(x)NIL R(x)1
EECE 411: Design of Distributed Software Applications
Sequential Consistency
The result of any execution is the same as if : operations by all processes on the data store were
executed in some sequential order, and the operations of each individual process appear in
this sequence in the order specified by its program.
Note: we talk about interleaved execution – there is some total ordering for all operations taken together
EECE 411: Design of Distributed Software Applications
Sequential Consistency (example)
Process1: Process 2 Process 3
X 1 Y 1 Z 1print (Y, Z) print (X, Z) print (X, Y)
Is output: 00 00 00 permitted?What about: 11 11 11?
What about: 00 10 10?
EECE 411: Design of Distributed Software Applications
Causal Consistency (1)
Writes that are potentially causally related must be seen by all processes in the same order. Concurrent writes may be seen in a different order by different processes.
Causally related relationship: A read is causally related to the write that provided the
data the read got. A write is causally related to a read that happened before
this write in the same process. If write1 read, and read write2, then write1 write2.
Concurrent <=> not causally related
EECE 411: Design of Distributed Software Applications
Causal Consistency (Example)
This sequence is allowed with a causally-consistent store, but not with sequentially or strictly consistent store.
Note: W1(x)a W2(x)b, but not W2 (x)b W1(x)c
EECE 411: Design of Distributed Software Applications
Causal Consistency: (More Examples)
A correct sequence of events in a causally-consistent store.
A violation of a causally-consistent store.
EECE 411: Design of Distributed Software Applications
Increasing granularity: Grouping Operations
Basic idea: You don’t care that reads and writes of a series of operations are immediately known to other processes. You just want the effect of the series itself to be known.
One Solution: Accesses to synchronization variables are
sequentially consistent. No access to a synchronization variable is
allowed to be performed until all previous writes have completed everywhere.
No read data access is allowed to be performed until all previous accesses to synchronization variables have been performed.
EECE 411: Design of Distributed Software Applications
Consistency models
Data centric: Assume a global, data-store view (i.e., across all clients) Continuous consistency
Limit the deviation between replicas Models based on uniform ordering of
operations Constraints on operation ordering at the data-store
level Client centric
Assume client-independentviews of the datastore
Constraints on operation ordering for each client independently
EECE 411: Design of Distributed Software Applications
Continuous Consistency
Observation: We can actually talk a about a degree of consistency:
replicas may differ in their numerical value replicas may differ in their relative staleness Replicas may differ with respect to (number and
order) of performed update operations
Conit: consitency unit specifies the data unit over which consistency is to be enforced.
EECE 411: Design of Distributed Software Applications
Example
Conit: contains the variables x and y: Each replica maintains a vector clock B sends A operation [<5,B>: x := x + 2]; A has made
this operation permanent (cannot be rolled back) A has three pending operations order deviation =
3
EECE 411: Design of Distributed Software Applications
Data centric: solutions at the data store level Continuous consistency
limit the deviation between replicas, or Models based on uniform ordering of
operations Constraints on operation ordering at the data-
store level
Client centric Assume client-independent views of the
datastore Constraints on operation ordering for each
client independently
Consistency models: contracts between the data store and the clients
EECE 411: Design of Distributed Software Applications
Eventual Consistency
Idea: If no updates take place for a long enough period time, all replicas will gradually (i.e., eventually) become consistent.
When? Situations where eventual consistency models may make sense
Mostly read-only workloads No concurrent updates
Advantages/Drawbacks
EECE 411: Design of Distributed Software Applications
Client-centric Consistency Models
System model Monotonic reads Monotonic writes Read-your-writes Write-follows-reads
Goal: Avoid system-wide consistency, by concentrating on what each client independently wants, (instead of what should be maintained by servers with a global view)
EECE 411: Design of Distributed Software Applications
Example: Consistency for Mobile Users
Example: Distributed database to which a user has access through her notebook.
Notebook acts as a front end to the database. At location A user accesses the database with
reads/updates At location B user continues work, but unless it
accesses the same server as the one at location A, she may detect inconsistencies:
updates at A may not have yet been propagated to B user may be reading newer entries than the ones available at
A: user updates at B may eventually conflict with those at A
Note: The only thing the user really needs is that the entries updated and/or read at A, are available at B the way she left them in A.
Idea: the database will appear to be consistent to the user
EECE 411: Design of Distributed Software Applications
Example - Basic Architecture
EECE 411: Design of Distributed Software Applications
Client-centric Consistency
Idea: Guarantees some degree of data access consistency for a single client/process point of view.
Notations: Xi[t] Version of data item x at time t at local
replica Li
WS(xi [t]) working set (all write operations) at Li up to time t
WS(xi [t]; xj [t]) indicates that it is known that
WS(xi [t]) is included in WS(xj [t])
EECE 411: Design of Distributed Software Applications
Monotonic-Read Consistency
Intuition: Client “sees” the same or newer version of data.
Definition: If a process reads the value of a data item x, any successive read operation on x by that process will always return that same or a more recent value.
EECE 411: Design of Distributed Software Applications
Monotonic reads – Examples
Automatically reading your personal calendar updates from different servers.
Monotonic Reads guarantees that the user sees always more recent updates, no matter from which server the reading takes place.
Reading (not modifying) incoming e-mail while you are on the move.
Each time you connect to a different e-mail server, that server fetches (at least) all the updates from the server you previously visited.
EECE 411: Design of Distributed Software Applications
Monotonic-Write Consistency
Intuition: A write happens on a replica only if it’s brought up to date with preceding write operations on same data (but possibly at different replicas)
Definition: A write operation by a process on a data item x is completed before any successive write operation on x by the same process.
WS
EECE 411: Design of Distributed Software Applications
Monotonic writes – Examples
Updating a program at server S2, and ensuring that all components on which compilation and linking depends, are also placed at S2.
Maintaining versions of replicated files in the correct order everywhere (propagate the previous version to the server where the newest version is installed).
EECE 411: Design of Distributed Software Applications
Read-Your-Writes Consistency
Intuition: All previous writes are always completed before any successive read
Definition: The effect of a write operation by a process on data item x, will always be seen by a successive read operation on x by the same process.
EECE 411: Design of Distributed Software Applications
Read-Your-Writes - Examples
Updating your Web page and guaranteeing that your Web browser shows the newest version instead of its cached copy.
Password database
EECE 411: Design of Distributed Software Applications
Writes-Follow-Reads Consistency
Intuition: Any successive write operation on x will be performed on a copy of x that is same or more recent than the last read.
Definition: A write operation by a process on a data item x following a previous read operation on x by the same process, is guaranteed to take place on the same or a more recent value of x that was read.
Examples: news groups
EECE 411: Design of Distributed Software Applications
Writes-Follow-Reads - Examples
See reactions to posted articles only if you have the original posting
a read “pulls in” the corresponding write operation.
EECE 411: Design of Distributed Software Applications
Rodamap
Updating replicas Consistency models
What is the semantic the system provides in case of updates
(luckily) applications do not always require strict consistency
Consistency protocols: How is this semantic implemented
Replica and content management How many replicas? Where to place them? When to get rid of them?
Redirection/Routing Which replica should clients use?
EECE 411: Design of Distributed Software Applications
Reminder: System view
Management system: Controls the allocated resources and aims to provide replication transparency
Consistency model: contract between the data store and the clients
EECE 411: Design of Distributed Software Applications
Next: implementation techniques
Updating replicas Consistency models (how to deal with updated data)
(luckily) applications do not always require strict consistency
Consistency protocols: Implementation issues How is a consistency model implemented
Replica and content management How many replicas? Where to place them? When to get rid of them?
Redirection/Routing Which replica should clients use?
EECE 411: Design of Distributed Software Applications
Reminder: Types of consistency models
Consistency model: contract between the data store and the clients
the data store specifies precisely what the results of read and write operations are in the presence of concurrency.
Data centric: solutions at the data store level Continuous consistency: limit the deviation between
replicas, or Constraints on operation ordering at the data-store level
Client centric Constraints on operation ordering for each client independently
EECE 411: Design of Distributed Software Applications
Today: Consistency protocols
Question: How does one design the protocols to implement the desired consistency model?
Data centric Constraints on operation ordering at the data-store level
Sequential consistency. Continuous consistency: limit the deviation between
replicas
Client centric Constraints on operation ordering for each client independently
EECE 411: Design of Distributed Software Applications
Reminder: Monotonic-Read Consistency
Definition: If a process reads the value of a data item x, any successive read operation on x by that process will always return that same or a more recent value.
Intuition: Client “sees” the same or a newer version of the data.
EECE 411: Design of Distributed Software Applications
Implementation of monotonic-read consistency
Sketch: Globally unique identifier for each write
operation Quiz-like question: explain how exactly the globally
unique identifier is (1) generated, and (2) used. Each client keeps track of:
Write IDs relevant to the read operations performed so far on various objects (ReadSet)
When a client launches a read Client sends the ReadSet to the replica server The server checks that all updates in the ReadSet
have been performed. [If necessary] Fetches unperformed updates
Returns the value
EECE 411: Design of Distributed Software Applications
More quiz-like questions
Sketch a Design for the consistency model
Monotonic-writes Read-your-writes Writes-follow-reads
Write pseudo-code
EECE 411: Design of Distributed Software Applications
Today: Consistency protocols
Question: How does one design the protocols to implement the desired consistency model?
Data centric Constraints on operation ordering at the data-store level
Sequential consistency, causal consistency, etc Continuous consistency: limit the deviation between
replicas
Client centric Constraints on operation ordering for each client independently
EECE 411: Design of Distributed Software Applications
Overview: Data-centric Consistency Protocols
EECE 411: Design of Distributed Software Applications
Overview: Data-centric Consistency Protocols
EECE 411: Design of Distributed Software Applications
Usage: distributed databases and file systems that require a high degree of fault tolerance. Replicas often placed on same LAN.
Issues: blocking vs. non-blocking
Primary-Based Protocols
EECE 411: Design of Distributed Software Applications
Primary-backup protocol with local writes
The primary migrates to the node that executes the write
Usage: Mobile computing in disconnected mode Idea: ship all relevant files to user before disconnecting, and
update later.
EECE 411: Design of Distributed Software Applications
Overview: Data-centric Consistency Protocols
EECE 411: Design of Distributed Software Applications
Active Replication (1/2)
Model: Remote operations/updates (and not the result of state changes) are replicated
Control replication (and not data replication) Requires total update ordering (either through a
sequencer or using Lamport’s protocol). Example: Replicated distributed objects Problems
Operations need to be sequenced Dealing with workflows
EECE 411: Design of Distributed Software Applications
Replicated-Write: Active Replication (2/2) Dealing with workflows in a general setting
EECE 411: Design of Distributed Software Applications
Overview: Data-centric Consistency Protocols
EECE 411: Design of Distributed Software Applications
Replicated writes: Quorums
Problem: some replicas may not be available all the time.
leads to low availability (for writes)
Solution: Quorum-based protocols: Ensure that each operation is carried on enough replicas (not on all)
a majority ‘vote’ is established before each operation distinguish a read quorum and a write quorum:
EECE 411: Design of Distributed Software Applications
Valid quorums
Must satisfy two basic rules A read quorum should “intersect” any prior write quorum at
>= 1 processes A write quorum should intersect any other write quorum
So, in a group of size N: Qr + Qw > N, and Qw + Qw > N results in Q>w > N/2
EECE 411: Design of Distributed Software Applications
Quorums: Mechanics for read and write
Setup: Replicas of data items have “versions”, Versions are numbered (timestamped) Timestamps must increase monotonically and include
a process id to break ties
Reads are easy: Send RPCs until Qr processes reply Then use the replica with the largest timestamp
(the most recently updated replica)
EECE 411: Design of Distributed Software Applications
Quorums: Mechanics for read and write (cont)
Writes are trickier: Need to determine version number Need to make sure all replicas are updated
atomically Consequence: need a ‘commit’ protocol to protect from
concurrent writes
EECE 411: Design of Distributed Software Applications
Quorum based protocols: Write algorithm
EECE 411: Design of Distributed Software Applications
Consistency protocols for continuous consistency
Question: How does one design the protocols to implement the desired consistency model?
Data centric Constraints on operation ordering at the data-store level
Sequential consistency. Continuous consistency: limit the deviation
between replicas
Client centric Constraints on operation ordering for each client independently
EECE 411: Design of Distributed Software Applications
Continuous Consistency
EECE 411: Design of Distributed Software Applications
Continuous Consistency (reminder)
Observation: We can talk a about a degree of consistency:
replicas may differ in their numerical value replicas may differ in their relative staleness There may differences with respect to (number
and order) of performed update operations
Conit: consistency unit specifies the data unit over which consistency is to be measured.
EECE 411: Design of Distributed Software Applications
Continuous Consistency: Bounding Numerical Errors (I)
Principle: consider a conit (data item) x and let weight(W(x)) denote the numerical change in its value after a write operation W.
[for simplicity] Assume that for all W(x), weight(W) > 0
W is initially forwarded to one of the N replicas: the origin(W).
TW[i, j] are the writes executed by server Si that originated from Sj
TW[i, j]=Σ {weight(W)| origin(W)=Sj & W in log(Si)}
Note: Actual value of x: vi : value of x at replica i
EECE 411: Design of Distributed Software Applications
Continuous Consistency: Bounding Numerical Errors (II)
Reminder: TW[i, j] writes executed by server Si originated at Sj
Problem: We need to ensure that v(t) − vi < di for every Si.
Approach: Let every server Sk maintain a view TWk[i, j] for what Sk believes is the value of TW[i, j].
This information can be piggybacked when an update is propagated.
Note: 0 ≤ TWk[i, j] ≤ TW[i, j] ≤ TW[j, j]
Solution: Sk sends operations from its log to Si when it sees that TWk[i, k] is getting too far from TW[k, k],
in particular, when TW[k, k] − TWk[i, k] > di/(N − 1).
Note: Staleness can be done analogously, by essentially keeping track of what has been seen last from Sk
EECE 411: Design of Distributed Software Applications
Issue 1. Dealing with data changes Consistency models
What is the semantic the system implements (luckily) applications do not always require strict
consistency Consistency protocols
How to implement the semantic agreed upon?
Issue 2. Replica management How many replicas? Where to place them? When to get rid of them?
Issue 3. Redirection/Routing Which replica should clients use?
Rodamap
EECE 411: Design of Distributed Software Applications
Quorums
Servers
Clients
State:
…State:
…State:
…
write A w
rite
A
write A
X
[Barbara Liskov]
EECE 411: Design of Distributed Software Applications
Quorums
Servers
Clients
State:
…State:
…State:
…A A
X
[Barbara Liskov]
EECE 411: Design of Distributed Software Applications
Quorums
Servers
Clients
State:
…State:
…A
write B
write B w
rite
B
X
State:
…A
X
[Barbara Liskov]
EECE 411: Design of Distributed Software Applications
C-A-P choose two
C
A P
Fox&Brewer“CAP Theorem”
consistency
Availability Partition-resilience
Claim: every distributed system is on one side of the triangle.
CA: available, and consistent, unless there is a partition.
AP: a reachable replica provides service even in a partition, but may be inconsistent.
CP: always consistent, even in a partition, but a reachable replica may deny service without agreement of the others (e.g., quorum).
EECE 411: Design of Distributed Software Applications
Issue 1. Dealing with data changes Consistency models
What is the semantic the system implements (luckily) applications do not always require strict
consistency Consistency protocols
How to implement the semantic agreed upon?
Issue 2. Replica management How many replicas? Where to place them? When to get rid of them?
Issue 3. Redirection/Routing Which replica should clients use?
Rodamap
EECE 411: Design of Distributed Software Applications
Reminder: System view
Management system: Controls the allocated resources and aims to provide replication transparency
Consistency model: contract between the data store and the clients
EECE 411: Design of Distributed Software Applications
Consistency model to have in mind for next slides: eventual consistency
Example applications: web content distribution – HTML, videos, etc
EECE 411: Design of Distributed Software Applications
Replica server placement
Problem: Figure out what the best K places are out of N possible locations.
Select best location out of N – j, j:0..K-1 for which the average distance to clients is minimal. Then choose the next best server.
(Note: The first chosen location minimizes the average distance to all clients.)
Computationally expensive. Select the k largest autonomous system and place a
server at the best-connected host. Computationally expensive.
Position nodes in a d-dimensional geometric space, where distance reflects latency. Identify the K regions with highest density and place a server in every one.
Computationally cheap.
EECE 411: Design of Distributed Software Applications
Content replication and placement
Model: We consider data objects (don't worry whether they contain data or code, or both)
Distinguish different processes: A process is capable of hosting a replica of an object or data:
Permanent replicas: Process/machine always having a replica
Server-initiated replica: Process that can dynamically host a replica on request of another server in the data store
Client-initiated replica: Process that can dynamically host a replica on request of a client (client caches)
EECE 411: Design of Distributed Software Applications
Content Replication: Server-Initiated Replicas
Keep track of access counts per file, aggregated by considering server closest to requesting clients
Number of accesses drops below threshold D drop file Number of accesses exceeds threshold R replicate file Number of access between D and R migrate file
EECE 411: Design of Distributed Software Applications
Client initiated replicas
Issue: What do I propagate when content is dynamic? State vs. operations
Propagate only notification/invalidation of update (often used for caches)
Transfer data from one copy to another (often for distributed databases)
Propagate the update operation to other copies (active replication)
Note: No single approach is the best, but depends highly on (1) available bandwidth and read-to-write ratio at replicas; (2) consistency model
EECE 411: Design of Distributed Software Applications
Propagating updates (1/3)
Issue Push-based Pull-based
State of server List of client replicas and caches None
Messages sentUpdate message (and possibly fetch update later)
Poll and update
Response time at client
Immediate (or fetch-update time) Fetch-update time
Pushing updates: server-initiated approach, in which the update is propagated regardless whether target asked for it.
Pulling updates: client-initiated approach, in which client requests to be updated.
EECE 411: Design of Distributed Software Applications
Propagating updates: Leases (2/3)
Observation: We can dynamically switch between pull and push using leases:
Lease: A contract in which the server promises to push updates to the client until the lease expires.
Technique: Make lease expiration time dependent on system's behavior (adaptive leases):
Age-based leases: An object that hasn't changed for a long time, will not change in the near future, so provide a long-lasting lease
Renewal-frequency based leases: The more often a client requests a specific object, the longer the expiration time for that client (for that object) will be
State-based leases: The more loaded a server is, the shorter the expiration times become
EECE 411: Design of Distributed Software Applications
Propagating updates: Epidemic protocols (3/3)
Problem so far: server scalability Server needs to provide updates to all system participants What if Internet-scale system (P2P)?
Epidemic protocols: Nodes periodically pair up and exchange state (or updates) Push and/or pull exchanges
Issues Generated traffic – mapping on physical topology When is an update distributed to everyone?
Probabilistic guarantees Freeriding
Use: server-less environments, volatile nodes mobile networks, P2P systems
EECE 411: Design of Distributed Software Applications
Issue 1. Dealing with data changes Consistency models
What is the semantic the system implements (luckily) applications do not always require strict
consistency Consistency protocols
How to implement the semantic agreed upon?
Issue 2. Replica management How many replicas? Where to place them? When to get rid of them?
Issue 3. Redirection/Routing Which replica should clients use?
Summary